47 lines
2.4 KiB
Python
47 lines
2.4 KiB
Python
|
"""Frozen DTW sentence point spans, independent of ordinary token intervals."""
|
||
|
|
from token_alignment import BOUNDARY, TEXT_TOKEN_LIMIT
|
||
|
|
|
||
|
|
POLICY = "dtw-token-sentences-v1;centisecond-points;first-last-point-span;exact-text;no-labels"
|
||
|
|
|
||
|
|
|
||
|
|
def aligned_segments(document, offset_ms, duration_ms):
|
||
|
|
from run import segments
|
||
|
|
originals = segments(document, offset_ms, duration_ms)
|
||
|
|
raw_rows = [row for row in document["transcription"] if row["text"].strip()]
|
||
|
|
result = []
|
||
|
|
previous = 0
|
||
|
|
for index, (row, original) in enumerate(zip(raw_rows, originals)):
|
||
|
|
tokens = row.get("tokens")
|
||
|
|
if not isinstance(tokens, list) or not 0 < len(tokens) <= 4096:
|
||
|
|
raise ValueError("Missing or oversized DTW tokens")
|
||
|
|
accepted = []
|
||
|
|
for token in tokens:
|
||
|
|
token_id, text = token.get("id"), token.get("text")
|
||
|
|
if type(token_id) is not int or token_id < 0 or not isinstance(text, str):
|
||
|
|
raise ValueError("Invalid DTW token identity")
|
||
|
|
if token_id >= TEXT_TOKEN_LIMIT:
|
||
|
|
if not text.startswith("[_") or not text.endswith("]"):
|
||
|
|
raise ValueError("Unrecognized special token")
|
||
|
|
continue
|
||
|
|
point = token.get("t_dtw")
|
||
|
|
if type(point) is not int or not 0 <= point * 10 <= duration_ms or point * 10 < previous:
|
||
|
|
raise ValueError("Missing, invalid or nonmonotonic DTW point")
|
||
|
|
previous = point * 10
|
||
|
|
accepted.append((text, previous))
|
||
|
|
if "".join(token[0] for token in accepted) != row["text"]:
|
||
|
|
raise ValueError("DTW text differs from original recognized segment")
|
||
|
|
phrase = []
|
||
|
|
for token_index, token in enumerate(accepted):
|
||
|
|
phrase.append(token)
|
||
|
|
if BOUNDARY.search(token[0]) or token_index == len(accepted) - 1:
|
||
|
|
text = "".join(part[0] for part in phrase)
|
||
|
|
if text.strip():
|
||
|
|
result.append({"start_ms": offset_ms + phrase[0][1],
|
||
|
|
"end_ms": offset_ms + phrase[-1][1], "text": text,
|
||
|
|
"original_segment_index": index,
|
||
|
|
"original_start_ms": original["start_ms"],
|
||
|
|
"original_end_ms": original["end_ms"]})
|
||
|
|
phrase = []
|
||
|
|
if len(result) > 10000:
|
||
|
|
raise ValueError("Oversized DTW passages")
|
||
|
|
return result
|